Glossary
What is synthetic data?
Synthetic data is generated rather than collected, and labs now use it heavily; this page covers where it replaces real data and where it can't.
Updated
Definition
Synthetic data is information produced by a computer program, such as a language model, a simulator or a statistical generator, rather than recorded from real people, transactions or events. It is made to resemble real data enough to train, test or share in its place.
Three kinds of generated data
The term covers several different things. Statisticians have long produced synthetic tables that keep the overall patterns of a sensitive file, such as averages and correlations, without copying any real row.
Simulated and model-written data
Engineers use simulators to produce sensor readings, driving scenes or physics data that would be slow or dangerous to record. And since large language models arrived, the biggest source is models writing text for other models: worked maths solutions, coding exercises, question-and-answer pairs and rewrites of web pages.
What it changes for real data
That third kind is what most people now mean by synthetic training data, and it changes the market for real data. If a lab can generate a million practice problems overnight, it will not pay much for ordinary ones. What it still needs, and still buys, is covered below, along with other ways labs source training material.
Synthetic data vs real data: what the research says
- of the data used to align Nemotron-4 340B was synthetic
- Over 98%
- of phi-4's pretraining mix was synthetic
- About 40%
- unweighted synthetic tokens in phi-4's pretraining
- About 400B
| Finding | Who reported it | What it means for owners of real data |
|---|---|---|
| Over 98% of the data used to align the Nemotron-4 340B models was synthetically generated | NVIDIA, Nemotron-4 340B technical report (2024) | Generic instruction and chat examples are now cheap to generate, so they fetch little |
| Synthetic data was the largest part of phi-4's pretraining, about 40% of the mix and roughly 400 billion unweighted tokens | Microsoft, phi-4 technical report (2024) | Reasoning practice material can be produced in bulk by the labs themselves |
| Models trained only on synthetic data did worse on knowledge-heavy tests and hallucinated more, so phi-4 kept filtered web data and licensed books | Microsoft, phi-4 technical report | Factual, specialist and recent knowledge still has to come from real sources |
| Clean, correct natural data was described as crucial for seeding synthetic data | Microsoft, phi-4 technical report | High-quality real material is the input synthetic pipelines are built on |
| Training on model-generated content without care caused irreversible defects, and rare cases in the original data disappeared | Shumailov et al., Nature (2024) | Real records of unusual events are what generated data loses first |
| The same authors expect data from genuine human interactions to become increasingly valuable | Shumailov et al., Nature (2024) | Records of real behaviour gain scarcity as generated text spreads online |
| Replacing real data with synthetic data tended towards collapse; keeping the real data and adding synthetic data alongside it avoided it | Gerstgrasser et al., arXiv (2024) | Synthetic data works as an addition to real data, not a replacement |
What the research adds up to
Read together, these sources point the same way. Labs generate data where correctness can be checked and volume matters, such as maths, code and instruction-following. They keep paying for real data where the model needs facts it cannot invent, rare cases, and evidence of how people and systems actually behave.
Where real data keeps its price
- Rare events: fraud cases, equipment failures, unusual medical presentations. Generators tend to smooth these away.
- Specialist knowledge that never reached the open web: internal manuals, expert notes, engineering decisions and the reasons behind them.
- Histories of real work, such as how a codebase changed over years or how support cases were resolved, which show cause and effect over time.
- Seed material: a lab building a synthetic pipeline for a field still needs correct examples from that field to start from.
An insurer's claims notes
Illustrative example, not a real deal: a regional insurer holds ten years of claims notes. A lab could generate plausible claims text on its own, but not the real mix of rare, disputed and fraudulent claims, or how each was settled. Those are the parts worth licensing, once personal details are removed.
Synthetic does not automatically mean anonymous
Synthetic data is often offered as a privacy-safe stand-in for personal records. Sometimes it is, but a generator trained on personal records can reproduce parts of them, so treat the label as a claim to test rather than a guarantee.
Test before sharing
If you hold synthetic data derived from customer files, check a sample for personal details and compare it with proper anonymisation before sharing it.
Where DataBroker fits, and its limits
We don't generate or sell synthetic data, and we have no figures on what labs pay for it. DataBroker teaches people to broker real company data, the kind the research above says generated data can't replace, and provides tools to value and prepare it. If you want a synthetic data supplier or generator, this is the wrong place.
Sources
4 sources · checked Nov 9, 2026
- Nemotron-4 340B Technical Report (arXiv 2406.11704)Over 98% of alignment data synthetically generated · Checked
- Phi-4 Technical Report (arXiv 2412.08905)Synthetic data as bulk of training, 40% mix, about 400B tokens, knowledge and hallucination limits, licensed books, natural seed data · Checked
- Shumailov et al., AI models collapse when trained on recursively generated data (Nature, 2024)Irreversible defects, disappearing tails, value of genuine human interaction data · Checked
- Gerstgrasser et al., Is Model Collapse Inevitable? (arXiv 2404.01413)Replacing real data leads to collapse; accumulating real and synthetic avoids it · Checked
Synthetic data FAQ
Is synthetic data as good as real data?
For some jobs, yes: labs report generating most of their alignment examples and large amounts of reasoning practice. For factual knowledge and rare cases, research and lab reports say models trained only on synthetic data do worse, so real data is still needed.
What is model collapse?
The loss of quality that happens when models are trained over and over on content produced by earlier models. A 2024 Nature paper found that rare cases from the original data disappear first. Keeping real data in the mix appears to prevent it.
Can you sell synthetic data?
Yes, some companies do, but buyers who can run a capable model can often generate similar data themselves. Synthetic data tends to sell when it comes from a simulator or process the buyer doesn't have.
Will synthetic data replace the need to buy training data?
Not on current evidence. The same lab reports that describe heavy synthetic use also describe keeping licensed and real sources for knowledge, and using real data to seed the generators.
Related
- AI dataset
An AI dataset is the material a model learns from: what it contains, how it is organised, and what buyers check before licensing one.
- Is selling data legal?
Is selling data legal? A plain look at when a sale is allowed, when registration applies, and when to call a lawyer first.
- How to sell data to AI companies
Sell data to AI companies only after you can prove where it came from, that you may license it, and why it isn't online already.
Free tool
Try the PII scanner
A PII scanner searches text for personally identifiable information, such as email addresses, phone numbers and government ID numbers, so it can be removed or protected before the text is shared.
- Free
- No sign-up
- Runs in your browser